How do I know whether pytorch is built with magma-cuda80?

When I run torch.svd(torch.rand(3,3).cuda()) I encountered an error:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
<ipython-input-6-9c860559505b> in <module>()
----> 1 torch.svd(torch.rand(3,3).cuda())

RuntimeError: No CUDA implementation of 'gesvd'. Install MAGMA and rebuild cutorch (http://icl.cs.utk.edu/magma/) at /data/wanggu/software/pytorch/torch/lib/THC/generic/THCTensorMathMagma.cu:280

However, I have installed magma-cuda80 through conda install -c soumith magma-cuda80 and rebuilt pytorch.
The error still exists. Could anyone help me?

What is the exact command you are using to rebuild pytorch? did you set CMAKE_PREFIX_PATH as specified here?

I set the cmake prefix to which conda,do I need to set it again when I rebuild pytorch?

cmake prefix path is not which conda, it is "$(dirname $(which conda))/../"

export CMAKE_PREFIX_PATH="$(dirname $(which conda))/../"

2 Likes

Thanks @smth, it’s fixed now. I misunderstood the “anaconda root directory”. I thought it was where conda located. :slight_smile: